Repro - Less is More: Geometric Unlearning for LLMs with Minimal Data Disclosure